The Qruise ML Physicist is a unique combination of highly advanced numeric algorithms, dealing with differential equation integration, optimization and more. The first step is to create an initial digital twin of the physical device.
The Qruise digital twin is the most sophisticated and detailed simulation available for the underlying physics of quantum setups, including not only the quantum components but a detailed simulation of the surrounding control electronics, lasers, AOM, AWG’s. – anything and everything which can affect the devices’ performance and error rates. They determine how to control the qubits (device) in an optimal way by using optimal control theory techniques based on the digital twin.
As the digital twin is initially not very accurate, the control parameters must be further optimized using the quantum device itself, using a process called closed-loop calibration. Qruise then take all the data observed during the experiment, and use it to improve the digital twin, in a process they call “model learning”. This leads to a more accurate digital twin (model), which leads to more reliable optimal control results, and the cycle continues until they achieve a highly reliable model of the system. A model which can now provide insight into the sources of the post-optimization residual error rates. This process is the culmination of decades of research at the intersection of machine learning, optimal control and quantum computing. On-top of the above many additional ML capabilities are added, such as Bayesian optimal experiment design, adversarial model validation, error-budget (“what-if”) determination and next-gen hardware co-optimization.
One might wonder, “What makes this unique?” With Qruise, machine learning is not simply throwing mountains of neural networks at the problem. It is a judicial and careful mixing of physical model simulations, machine-learning-inspired algorithms and tools, and some neural networks, where applicable. For example, they use reinforcement learning to tailor the various optimizers to the specific optimization problem (model-based optimal control, model-free calibration and model learning). Qruise employs Bayesian optimal experiment design to figure out what data they need to retrieve from the experiment in order to narrow the uncertainty of specific model parameters. Adversarial approaches are useful in testing the validity of models to the limits of the control hardware.